The method: how the receipt gets made.

Search first, judge second. The engine never decides whether a claim is true from memory; it retrieves real papers and reasons over what it actually found.

The pipeline, in plain English
Both sides of the questionYour claim becomes several search queries, aimed FOR it and AGAINST it.
Real papers onlyOpen academic indexes, citation counts attached, retracted papers binned on sight.
The evidence, weighedSupporting against contradicting, then the call: SUPPORTED, CONTESTED, REFUTED, or INSUFFICIENT EVIDENCE.
A score you can auditDeterministic confidence from source agreement, citation strength, replication, and counter-evidence.
Verbatim or nothingQuotes appear only word-for-word from the retrieved text. A missing quote never blocks a citation.
It corrects itselfA nightly scan re-checks every cited paper; a retraction downgrades the receipt publicly, on its own history.

Every receipt keeps an append-only history, and anyone can challenge one, but only with a source: opinions do not move verdicts, papers do. 41,593 claims are on the record so far.

The legend

Every stamp, and what it actually means.

A verdict is a claim about the evidence, not a claim about the truth. Here is the whole vocabulary, so nothing on a receipt has to be guessed at.

SUPPORTED

The retrieved evidence backs the claim and little or nothing credible contradicts it. It does not mean settled for all time; it means the record, as it stands today, is on this side.

CONTESTED

Credible evidence exists on both sides and we are not going to flatten it. A famous finding that failed replication lands here. This is the honest answer to a live dispute, not a failure to reach one.

REFUTED

The evidence weighs against the claim. Not "unproven": actively contradicted by what we found, with the contradicting sources attached.

INSUFFICIENT EVIDENCE

We looked and there is not enough to rule either way. We would rather hand you nothing than a citation that does not hold, and the miss is recorded publicly instead of hidden.

COMMON KNOWLEDGE

True by definition or everyday observation, and bolting a paper onto it would be noise dressed up as rigour. The engine refuses to cite "grass is green".

On a citation, you may also see
STRONG

The source says it explicitly. Cite it and move on.

DECENT

The source supports it with a short, fair inferential step.

WEAK

Related, but the leap is real. Use it knowing a marker might push back.

Those three grade how well ONE source carries ONE sentence, which is a different question from the verdict above. A meta-analysis can be STRONG with no verbatim quote in it, and a quote that cannot be matched word for word is dropped without dropping the citation.

And on a news claim, a LEAN instead
Some claims are too new or too journalistic for peer review: what an agency actually said, what a document actually records. Those get a lean rather than a verdict, and no model writes any part of it. We count what the evidence says on each side, weigh the record itself above anyone's opinion of it, let a publisher's newest ruling replace its own older one, and refuse to lean at all on a single voice. Out comes one of five:
strongly leans supportedleans supported evenly splitleans refutedstrongly leans refuted
Published accuracy →
We run a labeled evaluation suite (false-attribution rate, recall, quote fidelity, verdict accuracy) before every release, and we publish the numbers here; including the imperfect ones. Nobody else in this category shows you their accuracy. We think that's exactly the problem.